Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Standard view
Full view
of .
Look up keyword
Like this
0 of .
Results for:
No results containing your search query
P. 1
12-the choquet integral for the aggregation of interval scales in multicriteria decision making

12-the choquet integral for the aggregation of interval scales in multicriteria decision making



|Views: 43 |Likes:
Published by spitraberg

More info:

Published by: spitraberg on Feb 10, 2009
Copyright:Attribution Non-commercial


Read on Scribd mobile: iPhone, iPad and Android.
download as PDF or read online from Scribd
See more
See less


Fuzzy Sets and Systems 137 (2003) 11\u201326
The Choquet integral for the aggregation of interval scales in
multicriteria decision making\ue000
Christophe Labreuchea;\u2217, Michel Grabischb;1
aThales Research & Technology, Domaine de Corbeville, 91404 Orsay cedex, France
bUniversity of Paris VI, 4 place Jussieu, 75252 Paris, France

This paper addresses the question of which models \u00fft with information concerning the preferences of the decision maker over each attribute, and his preferences about aggregation of criteria (interacting criteria). We show that the conditions induced by these information plus some intuitive conditions lead to a unique possible aggregation operator: the Choquet integral.

\ue0002002 Elsevier Science B.V. All rights reserved.
Keywords:Multicriteria decision making; Choquet integral; Axiomatic approach
1. Introduction

Let us consider a decision problem that depends onn points of views described by the at- tributesX1;:::;Xn. The attributes can be any set. They can be of cardinal nature (for instance, the maximum speed of a car) or of ordinal nature (for instance, the color{red; blue; etc:}). We wish to model the preferences\u00a1of the decision maker (DM) over acts, that is to say over elements ofX =X1\u00d7 \u00b7\u00b7\u00b7 \u00d7Xn. For any x \u2208X, we use the following notation: x=(x1;:::;xn) wherex1\u2208X1;:::;xn\u2208 Xn. A classical way is to model\u00a1 with the help of an overall utility function

u: X\u2192R[11]:
\u2200x; y \u2208X; x\u00a1y \u21d4u(x)\u00bf u(y):
\ue000This paper is an extended and revised version of a paper presented in the EUROFUSE 2001 conference [12].
\u2217Corresponding author.
1On leave from Thales Research & Technology, Domaine de Corbeville, 91404 Orsay cedex, France.
0165-0114/03/$ - see front matterc
\ue0002002 Elsevier Science B.V. All rights reserved.
PII: S0165-0114(02)00429-3
C. Labreuche, M. Grabisch / Fuzzy Sets and Systems 137 (2003) 11\u201326
The overall utility functionu is often taken in the following way [11]:
u(x) = F(u1(x1);:::;un(xn));

where theui\u2019s are the so-calledutility functions, andF is the aggregation function. The main question is then how to determine theui\u2019s andF. The utility functions and the aggregation function are generally constructed in two separate steps. Each criterion is considered separately in the \u00ffrst step. The utility functionuirepresents the preferences of the DM over the attributeXi. The Macbeth approach [1,2] provides a methodology to construct theui\u2019s as scales of di\ue000erence (see Section6.1). In the second step, all criteria are considered together. As example of aggregation functions, one can \u00ffnd the weighted sum, the Choquet integral or the Sugeno integral. WhenF is a Choquet integral, several methods for the determination of the fuzzy measure are available. For instance, linear methods [15], quadratic methods [9,5] and heuristic-based methods [4] are available in the literature. For the Sugeno integral, heuristics [10] and methods based on fuzzy relations [16,17] can be found. Unfortunately, these methods address only the problem of constructing the aggregation function and refer more to learning procedures than to true decision making approaches. The way the \u00ffrst step is dealt with is generally not explained, especially with elaborate aggregators such as fuzzy integrals.

Yet, the two steps are intimately related. For instance, the notion of utility function has noabsolute meaning. One cannot construct theui\u2019s without any a priori knowledge of what kind ofF will be considered. However, when constructing theui\u2019s,F is not already known. As a consequence, the construction of the utility functions is more complicated than what it seems. One method addresses both points in a way that is satisfactory in the measurement theory standpoint: this is the Macbeth approach. However, it is restricted to the weighted sum. In such model, there is no interaction between criteria. The aim of this paper is to extend the Macbeth approach in such a way that interaction between criteria is allowed.

Theui\u2019s andF shall be determined from some information obtained from the DM about his preference over acts, e.g. elements ofX . The information we will consider in this paper can be seen as a generalization of the information needed in the Macbeth technology [1]. As in Macbeth, the information is based on the introduction of two absolute reference levels over each attribute, and the determination of scales of di\ue000erence, which, put together, ensures commensurateness. We will con- sider more general information than in the Macbeth approach in such a way that interaction between criteria will be allowed. We address then the question of which models \u00fft with this information. The information we have and the measurement conditions imply some conditions on the aggregation functionF. As an example, one condition is thatF shall enable the construction of theui\u2019s. Some conditions given on speci\u00ffc acts are naturally extended to wider sets of acts. These conditions rule out a wide range of families of aggregators. We see \u00ffnally which aggregation functions comes up. With these new properties, we show that the only possible model is the Choquet integral. We obtain an axiomatic representation of the Choquet integral that is a weak version of a result shown by Marichal [14].

We already addressed the same kind of problem for ratio scales in [7]. For ratio scales, the\u00c4

s integral seems to be the right aggregation function. The other di\ue000erence between this paper and [7] is that we restricted ourself only to the necessary information to construct the model (preferences over each attribute and the aggregation of criteria) and do not add any a priori assumption as in [7].

C. Labreuche, M. Grabisch / Fuzzy Sets and Systems 137 (2003) 11\u201326
2. Available information
The set of all criteria is denoted byN ={1;:::;n}. Considering two actsx; y\u2208 X andA\u2282 N , we
use the notation (xA;y\u2212A) to denote the actz\u2208 X such thatzi=xiifi\u2208 A andzi=yiotherwise.

It is well known in MCDM that the preference relation\u00a1can be modeled by an overall utility functionu (see (1)) if and only if all attributes are set commensurate in some way [11]. The notion of commensurateness hinges on the idea that we shall not try to aggregate directly attributes but rather aggregate values that represent the same kind of quantity. This is the commensurate scale.

The modeling of the preference relation of the DM depends entirely on what type of scale the DM can work on. Here, we assume that the DM can handle a scale of di\ue000erence. This means that, on top of being able to rank two actsx; y\u2208 X , the DM can also give an assessment of the di\ue000erenceu(x)\u2212u(y) between the overall utility ofx andy. In other words, ifx\u00a1y, the DM can give the intensity with whichx is preferred toy. We assume furthermore that the underlying scale is a bounded unipolar scale. It is bounded from above and below. In this case, the commensurate scale depicts the satisfaction degree of the DM over attributes. The satisfaction degree is typically a number belonging to the interval [0; 1]. The two bounds 0 and 1 have a special meaning. The 0 satisfaction value is the value that is considered completely unacceptable by the DM. The 1 satisfaction value is the value that is considered perfectly satisfactory by the DM. We assume that these two values 0 and 1 of the satisfaction scale can be identi\u00ffed by two particular elements0i and1i for each attributeXi. These two particular elements have an absolute meaning throughout the attributes. We assume that0iis the worst element ofXi, that is to say

Similarly,1iis the best element ofXi:
\u2200xi\u2208Xi;(xi;0\u2212i)\ue001 (1i;0\u2212i):
The introduction of0iand1ienables us to construct intra-criterion and inter-criteria information.
2.1. Intra-criterion information

For each criterioni, the mapping (denoted byui) from the attributeXito the satisfaction scale [0; 1] must be explicited. This corresponds to the preferences of the DM over each attribute. Com- mensurateness implies that the elements of one attribute shall be compared to the elements of any other scale. Taking a simple example involving two criteria (for instance consumption and maximal speed), this amounts to know whether the DM prefers a consumption of 5 l=100 km to a maximum speed of 200 km=h. This does not generally make sense to the DM, so that he is not generally able to make this comparison directly. In order to solve this problem out, the Macbeth approach is based on the idea that a scale of di\ue000erence is constructed separately on each attribute. A scale of di\ue000erence is given up to two degrees of freedom. Fixing two points on the scale determines entirely the scale. These two points are chosen in order to enforce the overall commensurateness. As a consequence, these two points are the only elements of the attribute that have to be compared to the elements of the other attributes. These elements are actually the0i\u2019s and1i\u2019s. All the0i\u2019s have the same meaning:u1(01) =\u00b7\u00b7\u00b7 =un(0n). Similarly,u1(11) =\u00b7\u00b7\u00b7 =un(1n).

You're Reading a Free Preview

/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->